OptiLIME

Formalizes the core tension in local explanations as a mathematical optimization problem. OptiLIME defines stability as the Pearson correlation coefficient of feature importance weights across multiple runs, and fidelity as the R-squared score of the surrogate model. It systematically traces the Pareto frontier between these two objectives to identify the optimal operating point.

Replaces ad-hoc manual tuning of the exponential kernel's sigma parameter with a deterministic algorithm. The framework evaluates candidate kernel widths by generating explanations at each setting and measuring the resulting stability-fidelity pair. It selects the largest kernel width that maintains a user-specified minimum stability threshold, maximizing local fidelity without sacrificing reproducibility.

Introduces a quantitative Stability Index to measure explanation robustness. For a given kernel width, multiple LIME explanations are generated with different random seeds. The index computes the average pairwise correlation of feature importance vectors. A value approaching 1.0 indicates near-identical explanations across runs, while lower values signal sensitivity to perturbation sampling noise.

Quantifies how well the sparse linear surrogate approximates the black-box model's decision boundary in the local neighborhood. OptiLIME uses the coefficient of determination (R²) on the weighted perturbed samples as the fidelity metric. A score of 0.95 means the surrogate captures 95% of the variance in the black-box's local predictions, ensuring the explanation faithfully represents model behavior.

Generates a Pareto frontier visualization plotting stability against fidelity for a range of kernel widths. This curve reveals the inherent trade-off: narrow kernels yield high-fidelity but unstable explanations, while wide kernels produce stable but lower-fidelity approximations. The optimal kernel width lies at the knee point of this curve, where marginal gains in one objective begin to severely penalize the other.

Operates as a wrapper around standard LIME, inheriting its model-agnostic property. OptiLIME can tune kernel widths for any data modality—tabular, text, or image—without requiring access to model internals. It only needs the black-box prediction function and a perturbation sampling strategy, making it deployable across diverse enterprise ML pipelines without architectural changes.

The hyperparameter directly optimized by OptiLIME. It controls the effective size of the local neighborhood by determining how quickly sample weights decay with distance from the instance.

• A narrow width enforces strict locality but can lead to high variance.
• A wide width smooths the explanation but may violate the local linearity assumption.
• OptiLIME finds the maximum width that preserves fidelity while maximizing stability.

The property that a local explanation remains consistent across multiple runs with different random seeds. OptiLIME directly targets this metric.

• Instability arises from the random perturbation sampling process.
• A stable explanation indicates that the identified features are robust signals rather than artifacts.
• OptiLIME defines stability as the concordance of feature rankings across repeated explanations.

A measure of how accurately the surrogate model approximates the black-box model in the immediate neighborhood. OptiLIME treats this as a constraint.

• Measured by the R-squared of the surrogate model on the weighted perturbed samples.
• OptiLIME enforces a minimum fidelity threshold to ensure the explanation remains meaningful.
• The framework explicitly balances fidelity against the kernel width to prevent over-smoothing.

The process of generating a synthetic neighborhood by randomly masking or altering features of the original instance. This randomness is the root cause of explanation instability.

• For text, this involves token masking.
• For images, this involves superpixel masking.
• OptiLIME's optimization directly counteracts the variance introduced by this sampling without requiring more samples.

The interpretable surrogate used within the LIME framework. OptiLIME optimizes the kernel width used to train this model.

• Uses Lasso regression to select only a few important features.
• The resulting coefficients become the feature importance scores.
• OptiLIME ensures these sparse weights are reproducible across different random seeds.

The fundamental balancing act that OptiLIME automates. A complex surrogate may be more faithful but less interpretable, while a simple one may lose precision.

• OptiLIME focuses on a sub-problem: the fidelity-stability trade-off.
• A kernel width that is too small yields high fidelity but low stability.
• The framework finds the Pareto-optimal point where stability is maximized without sacrificing fidelity below a set threshold.